Claude Opus 4.7 hits SOTA vision, engineering benchmarks

Opus 4.7 is a specialized "pro" upgrade that trades general-purpose intuition for elite technical performance and visual acuity. A 54.5% to 98.5% leap in visual performance makes it the first foundation model capable of handling dense engineering diagrams and high-resolution screenshots with production-grade reliability. The new "xhigh" effort level introduces a formal API tier for compute-intensive reasoning, allowing developers to pay more for deeper processing on complex tasks. Major regressions in NYT Connections and thematic reasoning suggest the model's weights have been aggressively optimized for logic and coding at the expense of "softer" intuitive benchmarks. Pricing parity with the previous generation ($5/1M input) signals Anthropic is aggressively defending its developer market share against OpenAI's GPT-5.4. Suspected "Adaptive" routing behaviors reported by users hint at the extreme compute challenges of serving high-effort models within a 1M token context window.